Remove empty string in the include and lib dir.

Ad dnn.include_dirs by default includa cuda_root, we remove it by default instead of doing a warning.

Move leftover flags

gpu dnn pool takes tensor variables

New problems cropped up on the buildbot

Add doc for opt.py

Don't let scan AllocEmpty cause false alarm by NanGuardMode

Try blas flags with rpath if they don't work.

tip for mixed dtypes

better stack trace handling.

Don't call eval

Fix crash with gemv and some broadcast pattern

cuDNN header / library paths fix

Fix the cnmem print percent message

Also lift Dot22Scalar.